Navigation
built from first principles on object storage: 10-100x cheaper, usage-based pricing, and massive scalability
╔═turbopuffer══════════════════════════════╗
╔════════════╗ ║ ║░
║ ║░ ║ ┏━━━━━━━━━━━━━━━┓ ┏━━━━━━━━━━━━━━┓ ║░
║ client ║░───API──▶║ ┃ Memory/ ┃────▶┃ Object ┃ ║░
║ ║░ ║ ┃ SSD Cache ┃ ┃ Storage (S3) ┃ ║░
╚════════════╝░ ║ ┗━━━━━━━━━━━━━━━┛ ┗━━━━━━━━━━━━━━┛ ║░
░░░░░░░░░░░░░░ ║ ║░
╚══════════════════════════════════════════╝░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
╔════════════╗
║ ║░
┌──║ client ║░
│ ║ ║░
API ╚════════════╝░
│ ░░░░░░░░░░░░░░
└─────────┐
│
▼
╔═turbopuffer══════════════════╗
║ ║░
║ ┏━━━━━━━━━━━━━━━━━━━━━━━━┓ ║░
║ ┃ Memory/ ┃ ║░
║ ┃ SSD Cache ┃ ║░
║ ┗━━━━━━━━━━━━━━━━━━━━━━━━┛ ║░
║ │ ║░
║ ▼ ║░
║ ┏━━━━━━━━━━━━━━━━━━━┓ ║░
║ ┃Object storage (S3)┃ ║░
║ ┗━━━━━━━━━━━━━━━━━━━┛ ║░
║ ║░
╚══════════════════════════════╝░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░Warm queries have all their data in cache.
Approach: Search for 100 random vectors from the 768d dataset with top k = 10, when dataset is fully in cache. Warming the cache for an index takes about ~10s (for 1m vectors) after the first cold query, and typically stays in cache for a few hours.
Aman Sanger
Co-founder
10B+
documents
5M+
namespaces
10×
cheaper
Justin Watts
Distinguished Engineer
Moving to turbopuffer felt less like an upgrade and more like discovering a new paradigm. We didn't just save costs; we turned once-prohibitive features into standard tools in our arsenal. Vectorize all the things and say goodbye to sharding
MTL 🇨🇦
single-tenant cluster
| Metric | Max seen in production | Production limits (current) | Production limits (soon) |
|---|---|---|---|
| Max documents (global) | 35B+ | Unlimited | |
| Max documents (per namespace) | 180M | 100M | 1B+ |
| Number of namespaces | 10M+ | Unlimited | |
| Max dimensions | 10,752 | ||
| Max inactive time in cache | ~3 days | Contact us for custom | |
| Write rate (global) | 100,000 doc/s | Unlimited | |
| Write rate (per namespace) | 10,000 doc/s | Unlimited | |
| Queries (global) | 1000+ queries/s | Unlimited | |
| Max queries/second (per namespace) | 100+ queries/s | 100+ queries/s | 10,000 QPS |
| Vector search recall@10 | ~90-95% | ~90-95% | Configurable |
| Max attribute value | 8 MiB | 8 MiB | |
| See full list and workloads that are a great fit for turbopuffer 🐡 |
FAQ
Yes, turbopuffer is production quality. We have powered production applications since November '23 at 99.99% uptime. We host billions of production vectors at thousands of writes per second.
We maintain SOC2 Type 2 certification and HIPAA compliance.
turbopuffer accounts are not yet generally available. We are currently focused on providing high bandwidth support to make our select early customers successful, while removing limitations. See if your workload is a good fit for turbopuffer here.
